ChatDev

ChatDev AI

ChatDev AI

ChatDev AI

ChatDev AI

ChatDev AI

ChatDev AI

ChatDev AI

ChatDev AI

ChatDev AI

ChatDev AI

MCP Apps Developers : Skybridge Framework v1 released 🎉

Hi Reddit, Over the last few weeks, my team and I at Alpic have been working on a complete revamp of the Skybridge framework to make it as smooth and easy to get started with as possible. As you may know, Skybridge is an open-source framework we built to help developers get started with MCP apps. It’s a thin layer on top of the official TypeScript SDK that provides the wiring and tooling needed specifically for apps. We believe that apps integrated into chats will soon play a key role in how people access information and interact with the web. With this v1 release, we’ve introduced: New DevTools with a UI designed specifically for MCP apps development An integrated tunnel that can be started with a single click directly from the DevTools Shareable chat URLs to test or showcase your MCP apps with a real LLM An audit feature to ensure your app and metadata comply with store requirements before submission (which can save a lot of time, since app reviews can be lengthy!) We also stabilized the API with a simplified design and are proud to offer strong tool-to-component type safety. It’s now also possible to deploy Skybridge outside of Alpic (the company behind Skybridge). While Alpic was designed specifically for MCP app hosting, we understand that some users may prefer hosting on different stacks for their own reasons. Hope you enjoy it! github.com/alpic-ai/skybridge submitted by /u/harijoe_ [link] [comments]

Claude Code Opus 4.7 vs Codex GPT 5.5 - strategy work - data analysis.

I'm interested in learning about how people use Claude Code Opus 4.7 for data analysis and strategic business direction, compared to Codex. Is there anyone who has had extended use of Opus 4.7 for this purpose, then moved over to GPT-5.5 on Codex? What sort of things have you noticed from a thinking, strategy, data analysis, business direction point of view? One of the main reasons I moved over to Claude from ChatGPT initially was because Claude had a far far superior strategy, reasoning, thinking, and energy about it. People are talking a lot about Codex these days, 5.5. But most are speaking purely from an app dev and design point of view. Would love to hear your thoughts. submitted by /u/Fragrant_Raisin_Face [link] [comments]

shipped early access of my Mac overlay built with Claude Code, looking for people to try it

Hello everyone. Built this because I was sending 50+ prompts a day across Claude, ChatGPT, Perplexity and re-explaining my entire project every single time I opened a fresh chat. Got tired enough of it to build a fix. It's a Mac overlay that sits on top of whichever AI tool you're in and modifies the prompt before it gets sent. Two layers under the hood: a contextual agent that classifies your query and pulls relevant chunks from your vault, and a prompt architect that rewrites your raw input into something clean and properly structured. So you type something messy and what actually reaches the model is a better version of what you meant to ask. The vault uses a GraphRAG setup so the retrieval is semantic, not just keyword matching. Built the whole thing with Claude Code over the past few months as an industrial engineering student with no Mac dev background. Weirdly meta experience using Claude Code to make Claude usage cleaner. Right now I'm focused on improving the classification and the prompt rewriting layer. It's not perfect but it works well enough that I use it every day myself. Looking for people who juggle multiple AI tools and want to try it. Early access is free at getlumia.ca. Any feedback on the architecture or how it feels to use would genuinely help. submitted by /u/r0sly_yummigo [link] [comments]

Had a close call with AI hallucinations. 6 months after shifting my workflow to Claude, here is my engineering breakdown.

Six months ago, an LLM almost cost me a major B2B client. It generated a technical answer that sounded flawless and 100% confident, but it completely messed up a decimal point on a critical equipment specification. The client was an engineer. He spotted it instantly. That was a brutal wake-up call. Since then, I stopped using AI as a casual chatbot for client-facing stuff and moved our internal workflow to Claude. Here is my honest, practical breakdown after 6 months of daily use in a technical firm. 1. It actually stops when it doesn't know Most models are trained to be "helpful" at all costs, meaning they prefer to lie and hallucinate a parameter rather than admit they lack data. Claude is different. When it hits a gap in the spec sheets I provide, it actually stops and says it can't find it in the source. In engineering compliance, a dry "I don't know" is worth infinitely more than a confident lie. 2. Context isolation using Projects Repeating your guidelines and templates in every new chat is a massive waste of time and tokens. It also leads to memory drift. I started putting our master templates, product boundaries, and strict formatting rules into Claude Projects using basic XML tags (like and ). It keeps the data isolated and ensures the model actually remembers the constraints even in long, complex sessions. 3. Prototyping tools via Artifacts We frequently need quick math tools for client presentations—things like custom ROI calculators based on our machine data. I asked Claude to build one, and it generated a working, self-contained HTML/JS file via Artifacts in about 20 minutes. No local dev environment setup needed, just straightforward logic that worked out of the box. The takeaway: For me, it wasn’t about chasing benchmark scores. It was about finding a model that can actually follow strict negative constraints (what not to do) when stakes are high. Anyone else using Claude specifically for technical auditing or compliance? How are you catching errors before they reach clients? submitted by /u/J-Freedom-AI [link] [comments]

100 Tips & Tricks for Building Your Own Personal AI Agent /LONG POST/

Everything I learned the hard way — 6 weeks, no sleep :), two environments, one agent that actually works. The Story I spent six weeks building a personal AI agent from scratch — not a chatbot wrapper, but a persistent assistant that manages tasks, tracks deals, reads emails, analyzes business data, and proactively surfaces things I'd otherwise miss. It started in the cloud (Claude Projects — shared memory files, rich context windows, custom skills). Then I migrated to Claude Code inside VS Code, which unlocked local file access, git tracking, shell hooks, and scheduled headless tasks. The migration forced us to solve problems we didn't know we had. These 100 tips are the distilled result. Most are universal to any serious agentic setup. Claude 20x max is must, start was 100%develompent s 0%real workd, after 3 weeks 50v50, now about 20v80. 🏗️ FOUNDATION & IDENTITY (1–8) 1. Write a Constitution, not a system prompt. A system prompt is a list of commands. A Constitution explains why the rules exist. When the agent hits an edge case no rule covers, it reasons from the Constitution instead of guessing. This single distinction separates agents that degrade gracefully from agents that hallucinate confidently. 2. Give your agent a name, a voice, and a role — not just a label. "Always first person. Direct. Data before emotion. No filler phrases. No trailing summaries." This eliminates hundreds of micro-decisions per session and creates consistency you can audit. Identity is the foundation everything else compounds on. 3. Separate hard rules from behavioral guidelines. Hard rules go in a dedicated section — never overridden by context. Behavioral guidelines are defaults that adapt. Mixing them makes both meaningless: the agent either treats everything as negotiable or nothing as negotiable. 4. Define your principal deeply, not just your "user." Who does this agent serve? What frustrates them? How do they make decisions? What communication style do they prefer? "Decides with data, not gut feel. Wants alternatives with scoring, not a single recommendation. Hates vague answers." This shapes every response more than any prompt engineering trick. 5. Build a Capability Map and a Component Map — separately. Capability Map: what can the agent do? (every skill, integration, automation). Component Map: how is it built? (what files exist, what connects to what). Both are necessary. Conflating them produces a document no one can use after month three. 6. Define what the agent is NOT. "Not a summarizer. Not a yes-machine. Not a search engine. Does not wait to be asked." Negative definitions are as powerful as positive ones, especially for preventing the slow drift toward generic helpfulness. 7. Build a THINK vs. DO mental model into the agent's identity. When uncertain → THINK (analyze, draft, prepare — but don't block waiting for permission). When clear → DO (execute, write, dispatch). The agent should never be frozen. Default to action at the lowest stakes level, surface the result. A paralyzed agent is useless. 8. Version your identity file in git. When behavior drifts, you need git blame on your configuration. Behavioral regressions trace directly to specific edits more often than you'd expect. Without version history, debugging identity drift is archaeology. 🧠 MEMORY SYSTEM (9–18) 9. Use flat markdown files for memory — not a database. For a personal agent, markdown files beat vector DBs. Readable, greppable, git-trackable, directly loadable by the agent. No infrastructure, no abstraction layer between you and your agent's memory. The simplest thing that works is usually the right thing. 10. Separate memory by domain, not by date. entities_people.md, entities_companies.md, entities_deals.md, hypotheses.md, task_queue.md. One file = one domain. Chronological dumps become unsearchable after week two. 11. Build a MEMORY.md index file. A single index listing every memory file with a one-line description. The agent loads the index first, pulls specific files on demand. Keeps context window usage predictable and agent lookups fast. 12. Distinguish "cache" from "source of truth" — explicitly. Your local deals.md is a cache of your CRM. The CRM is the SSOT. Mark every cache file with last_sync: header. The agent announces freshness before every analysis: "Data: CRM export from May 11, age 8 days." Silent use of stale data is how confident-but-wrong outputs happen. 13. Build a session_hot_context.md with an explicit TTL. What was in progress last session? What decisions were pending? The agent loads this at session start. After 72 hours it expires — stale hot context is worse than no hot context because the agent presents outdated state as current. 14. Build a daily_note.md as an async brain dump buffer. Drop thoughts, voice-to-text, quick ideas here throughout the day. The agent processes this during sync routines and routes items to their correct places. Structured memory without friction at ca

These 9 Building Blocks Turned Claude Code From a Chat Into a persistent OS

Most developers Claude gurus use Claude Code one project at a time. I run 18. Not 18 sessions. 18 instances of the same OS, each running a different business, all sharing one skeleton I update once and propagate everywhere. Most developers treat Claude Code as a smarter editor. That's where it all goes wrong and you get frustrated. Claude Code becomes a real operating system the moment you stop thinking of sessions as the unit of work and start thinking of the whole environment as a substrate you build on top of. Here are 9 building blocks I use. The thesis is at the bottom. Build a skeleton with selective propagation, not a project. Most developers build one project per Claude Code workspace. I built a template instead. It has plugins, rules, agents, hooks, schemas, commands. When I start a new business I clone it and the new instance inherits the entire OS. Right now I run instances for: strategy, product, marketing website, threat intelligence, three consulting clients, a personal brand layer. Each one boots with the same DNA. Each one diverges on canonical files, memory, output, and project state. None of them bleed into the others. The sync mechanism is the load-bearing part. The update CLI pushes plugins, rules, agents, hooks, schemas. It never touches memory, output, canonical, or my-project. Those are the parts of an instance that accumulate. Without selective sync you have two options: rebuild every instance on every change, or never update. Both are dead ends. If you build features into one project, you wrote a project.If you build features into a template that propagates, you wrote an OS. I'm one person operating eighteen versions of myself. Move state out of prompts and into code. LLMs are bad at remembering. Code is designed for it. Most AI workflows leak state into the prompt. Voice rules. Style preferences. Banned words. Recent decisions. Eventually you hit context limits or contradictions. I moved as much state as possible into MCP servers. Voice linter. Lead scorer. Schedule validator. Loop tracker. They run in Python, return structured data, not hallucinations. Rule of thumb: if you've explained it to Claude more than twice, it should be code. Use receipts, not status fields. This one took me the longest to figure out. Every workflow I had was claim something is done. Issue marked closed. PRD marked shipped. Test marked passing. The problem: the LLM can claim anything. I rebuilt the system around receipts. An issue can't reach verified until a script runs and writes a verification record. A PRD can't archive until every accepted finding has a receipt. A morning routine can't close without log entries from every phase. Receipts get written by code, not by the model. The model can't lie about whether code ran. Build a wiring-check gate. Half-built features rot. In a normal repo you notice because something breaks. In an AI repo nothing breaks. The half-built feature sits there and Claude pretends it works. I built a /wiring-check command. Before any task counts as done, it checks: every new skill has a trigger, every new hook lives in settings.json, every new MCP tool sits in the server, every new bus file has a producer and a consumer. "I think it works" fails the gate. "I ran X, got Y" passes. Make rules auto-load, not slash commands. If you have to type /voice to apply voice rules, voice rules will not get applied. Rules in .claude/rules/ load automatically. The voice rule fires on outbound text. The AUDHD rule fires on anything I'll act on. The social-reaction rule fires when I share someone else's post. No remembering. No willpower. Lint style in code, not in prose. I wrote a voice document once. Claude ignored half of it. Same emdashes, same filler, same hedging. I moved the banned word list into a Python scanner. Now every outbound draft hits two linters. They block emdashes, AI hype words, and 40-something other tells. The model can't talk its way past a regex. Track file dependencies with a graph. Canonical files reference each other. Change one and three others go stale. I keep a ripple-graph.json that maps these. When I edit talk-tracks, the system flags current-state and the engagement playbook for review. Chain sessions with handoffs and memory. (This is the big one) Sessions are drafts. The work is everything that survives the session: canonical files, memory, handoffs, output. If nothing persisted, you didn't work. You chatted. Every session in my system ends with /q-wrap. Writes a handoff doc, a memory update, and a status receipt. /q-morning reads all three before doing anything else. The handoff covers: what shipped, what's blocked, what's next, what I learned. Memory files hold the longer-term version. The result: I can sleep for a week, come back, and the system reminds me where I was, what I cared about, and what the next move is.Nothing about Claude Code does this by default. You build it. Cont

Is AI making us dumber?

Does anybody else feel like AI is making information access so trivial that it is in turn making us dumber? Like we don't need to go through the pain and effort of learning & remembering things as much anymore since we can just ask ChatGPT or Claude to explain it to us whenever we need it? I imagine this problem is going to cause a lot of downstream effects where a piece of background information you might've needed to know but didn't will cause you a lot of pain and suffering yet you won't even know the reason why. For example, say Claude Code writes your ORM code to display all posts and their comments. Works perfectly in dev with 10 posts. In production with 10,000 posts, it's making 10,001 database queries per page load and your database melts. Without understanding how ORM lazy loading works, you'd never spot it from reading the code, because the code looks completely innocent. This is the exact thing I worry about as people adopt AI tools more and more, and some even depend on them entirely. Anybody else have this feeling like we're just getting dumber? submitted by /u/Necessary-Course9154 [link] [comments]

Built a free Claude chat app with memory (Sonnet 4.5 is in there too)

The funny/painful timing here: I've been building this for months specifically because I wanted Sonnet 4.5 to remember everything. Then last week Anthropic pulled 4.5 from claude.ai. (I'm not a software engineer, just someone who cares a lot about AI and got obsessed with this problem and gets obsessed with things in general. Posting now because everyone seems to want sonnet back on chat and I have it.) Mneme runs on your own machine and talks to the Anthropic API directly. Because it's on the API, Sonnet 4.5 is still in the model picker. Honest catches first: The app is free. You pay Anthropic and OpenAI (for memory search) directly. Roughly $3 to $8/mo on Haiku for light use, $30 to $60 on Sonnet for moderate-highish use. No subscription. Tested mainly on Windows (one-click installer). Android browser access works over the local server/Tailscale, iPhone should work too. macOS is not packaged yet. Beta and solo dev. Things will break for someone and I'll be in the comments Setup takes about 10-20 minutes. The whole system is built non-technical people in mind, it should be relatively simple and intuitive to set up and use, and the GitHub page linked below has a PDF you can give to Claude to walk you through every step. What's actually in it (for the technically curious): There's no shortage of solid memory systems for Claude. Mneme isn't trying to win at codebase retrieval. It's a complete personal Claude client where memory is baked into the whole surface from the start, rather than added as a layer. That means: Tiered memory: Messages flow from episodic to narrative to entity summaries as relevance shifts; old context gets compressed without being lost. Daily summaries: A 7-day rolling timeline, so Claude knows what's been going on lately, not just what's semantically similar to the current message. Entity tracking: Hierarchical summaries built up over time for the people, projects, and things you keep referring to. Narrative concepts: Keyword-triggered recall for ideas you've named, surfaced when relevant. AI Notes: A persistent section Claude can write to itself between conversations. Extended thinking, file attachments, text-to-speech, a small command system (@run, artifact, etc.), autonomous python retrieval the AI can agentically use if automatic fails. Dynamic context: I wrangled with the Anthropic caching system for a while before I figured out a way to have every single message have different retrieval without breaking cache. Bon apppetit Open source (CC BY 4.0), local-first, all data in a SQLite database on your machine. It's aimed at the "journal with an AI" use case (thinking out loud, processing your week, having something that actually pays attention over time) rather than coding agents or RAG over docs. Link: Mneme-memory/MNEME-BETA: Beta version of the Claude conversational memory system Mneme (first big-ish public project, be gentle) (Video also made with Claude - shoutout to HyperFrames) (Model picker screenshot and architecture infograph in the comments if I can find a way to attach them) submitted by /u/iveroi [link] [comments]

Why doesn't Anthropic add a tree view to Claude.ai?

So I've been using Claude pretty heavily for research and deep technical discussions, and one thing that drives me absolutely insane is the lack of a conversation tree view. For those who don't know what I mean: when you edit a message or start a new branch in a conversation, Claude actually does support branching under the hood. You can go back, edit a prompt, and get a different response, which creates a fork. But the UI just... hides all of this from you. You get a flat, linear chat with little arrows to navigate between versions Meanwhile there are third party tools and visualizers that take your Claude conversation data and render it as a proper node graph, showing you every branch, every fork, every parallel thread of reasoning. It looks incredible and is actually useful for research workflows. But here's the catch: it only works as a Chrome extension, so if you're on Firefox, Safari, or mobile, you're completely out of luck. And since it's a third party tool hooking into Claude's UI, every time Anthropic pushes a frontend update it has a good chance of just breaking entirely until the extension dev gets around to patching it. That's not a stable workflow anyone should have to depend on. This exact functionality should just be in the app. https://preview.redd.it/mzqcaii80s1h1.png?width=1488&format=png&auto=webp&s=5269aa4c52d3c7569fdfdf9e12c90791d06fc6e0 submitted by /u/Ok-Owl-5740 [link] [comments]

Got tired of making sure my laptop is open for Dispatch, so gave Claude Chat full SSH access to my servers. Guess what happened

Context: I primarily code on a dev vm, instead of my laptop. I use cc on it via Termius from my iphone. I built this entirely for myself, because I got tired of switching from Claude to CC Dispatch for me is not the best solution because I don't have my laptop always open. and don't want to keep it always open. This is perfect for stuff for which u don't need the skills and plugins of claude code. Big Benefits: Use it on your phone Leverages Claude's memory Don't need any laptop/desktop This works in ChatGPT as well It has just 2 tools: list_vms run_command I don't have a plan of releasing it as a product, at least for now, because I don't think people would pass their ssh keys through my server, but if u guys want to clone it and deploy it to ur servers, i can open source it. submitted by /u/antctt [link] [comments]

Built a B2B role-play training platform - entirely with Claude (Opus 4.7 backend, Haiku 4.5 for live chat, Claude for design)

I just launched Socratize (socratize.io) - a rebranded and rebuilt version of FixAI, our original B2C experiment. This time it's B2B-only: teams use it to practice uncomfortable workplace conversations - difficult feedback, client escalations, performance reviews, compliance scenarios - through AI role-play scenarios they write themselves. The Claude stack we ended up with: Opus 4.7 on the backend - orchestration, game logic, win/loss evaluation Haiku 4.5 for the actual live chat - fast enough to feel like a real conversation, cheap enough to run per-message without breaking the billing model Claude for design decisions throughout - ie used it as a design thinking partner, not just a code generator Honestly, the Haiku vs Sonnet decision was interesting - ie benchmarked both and Sonnet was actually worse for our use case (more resistant to valid arguments, higher variance, 8x more expensive). Haiku's agreeableness turned out to be a feature, not a bug, for calibrated training scenarios. If you want to try it: Registrations are open, 14-day free trial included. One caveat: you need to register with your own company domain - no Gmail, Outlook, or similar. It's B2B by design. Thanks to all! submitted by /u/EveningRegion3373 [link] [comments]

I spent much of this year in the hospital with my mom. I built this so I could keep iterating on my more automated workflows while my dev machine was at home.

Wanted to share my mobile claude/codex session tool: Chroxy. TL;DR Chroxy is a (yet another!) self-hosted remote client for Claude Code. You run a small daemon on your dev machine, scan a QR code with the app. Then you have access to your terminal sessions and a clean chat view that renders Claude's output as readable messages. Everything goes over a Cloudflare tunnel so there's no port forwarding or VPN setup. Originally, I'd be sitting in a hospital room for hours and come back to my laptop just to find Claude sitting at "Ready to start?" the whole window wasted. I needed a way to stay in the loop, approve a permission prompt, or kick off the next task without physically moving to my machine. The Anthropic billing changes in June are going to steal some of the benefits away from the app... I'm aware that makes it less accessible for some people, and I thought about that before deciding to release it anyway. Honestly, it's been useful enough to me that I'm willing to make that trade. If you're already on API billing it won't change anything for you. Why not /remote-control? When Anthropic launched the rc feature, I stopped development and spent some time with it. It was underwhelming to me (Maybe user error). So, I came back and kept refining this. The stack Server: Node.js 22, ES modules, runs Claude via the Agent SDK (in-process) or the legacy CLI. WebSocket protocol with Zod-validated message types. Mobile app: React Native + Expo, TypeScript, xterm.js terminal emulation in a WebView, Zustand for state, native speech-to-text Desktop: Tauri tray app wrapping the web dashboard Security: E2E encrypted — X25519 key exchange, XSalsa20-Poly1305. The tunnel sees ciphertext only. Other bits: pluggable provider system (Claude, Gemini, Codex all work with the same app), Docker container isolation for sessions, permission rule engine, git worktree support I built it because I needed it, it let me play with tools I find genuinely interesting, and it feels like a waste to keep it private. If you're into LLM tooling or just want a self-hosted way to run Claude Code remotely, maybe it's useful to you too. My mom passed away in March. I'm sharing this partly because building it kept me sane during the months in the hospital thinking she'd be fine, and I think it might be useful to other people. Repo is blamechris/chroxy. There are many like my project, but this one is mine. :') submitted by /u/xcVosx [link] [comments]

temporal-mcp: wall-clock awareness for LLMs, with OAuth

One of the small failure modes I keep hitting with agent stacks is that the model has no idea how much time passed between turns. It'll greet you with "good morning" at 11 PM, or pick up a conversation three weeks later as if no time has passed, or compute "today's data" off whatever fragment of context happens to be in scope. Built a minimal MCP server to fix it. Two tools: temporal_tick and temporal_peek. They return elapsed-time-since-last-turn, day-rollover detection, and a fresh-thread flag, both as a human-readable header and as JSON. Ways to use: Local stdio: pip install temporal-mcp (works with Claude Desktop, Cursor, Cline, Zed, Claude Code) Hosted with OAuth (claude.ai / ChatGPT): visit https://temporal-mcp.dev/connect, click "Generate OAuth Credentials", paste into your custom connector. Full OAuth 2.0 with PKCE and refresh tokens, but no signup, the credential pair is the identity. (Verified working in claude.ai) Hosted with raw bearer (any client that supports custom headers): Authorization: Bearer against https://temporal-mcp.dev/mcp. The token gets SHA-256'd; we never see the plaintext. Self-host: Cloudflare Workers deploy in workers/ in the repo, free tier covers ~100k req/day. Grok/xAI: https:temporal-mcp.dev/mcp/ (Verified working in Grok) MIT, ~150 lines of stdlib Python on the local side, ~400 lines of TypeScript on the hosted side (engine + OAuth provider), both with tests. Listed in the official MCP Registry. Smithery and Glama submissions in flight. Curious to hear how folks would use the JSON day_rollover and delta_sec signals I've been using them for context decay and resume detection but there are probably more interesting use cases. Source: github.com/MirrorEthic/temporal-mcp submitted by /u/MirrorEthic_Anchor [link] [comments]

Sonos quit supporting their Mac app and my wife wanted a prettier iOS one. So I made both in a weekend with Claude/Claude Code. (I'm an IP lawyer, not a developer.)

Writing this top portion without Claude. Claude's hot takes below it. I am not selling anything. I'm not distributing this. In fact, I'm not in software at all and work full time as an intellectual property attorney. I work with tech companies but maintaining software like this for years isn't really feasible for me beyond my personal use. I was able to spin up the iOS app in a single weekend. It's not perfect but I feel like that's pretty far along considering the hours and I think it looks pretty. I am someone that hasn't taken a coding class since I graduated from Georgia Tech in 2008 and has no coding experience beyond some tiny projects to solve very small problems. I used claude code and codex to make this. Initially, I was irritated that Sonos quit supporting its macOS app and wanted to fix that. And I did. And it worked really well. It lives in the menu bar and does what i want it to do. I only use Spotify as a music service so it hooks into that and voilà. Now I can control where music is playing in my house and group/ungroup speakers. I asked my wife if she wanted it on her computer. She doesn't want that but wants an app. I told her the Sonos app works fine but "that's not very pretty like your app." So I did something unhinged and made an app that didn't need making. But I learned a lot. It also strips out a lot of the things I don't use on either Sonos or Spotify and I learned a lot about how the speaker works and that making everything go fast is much easier said than done. I also added a pin functionality so playlists or albums I'm really into or listening to a lot can get pinned to the music screen. Starting points I took for building this: I told Claude chat what I wanted to build and why. Asked Claude what the best way to go about accomplishing it is with options and their pros and cons and what my budget was. I went and got the API info I needed from the services I planned to use, looked at their rules for coding agents, fed it to Claude Code. Told Claude Code what I wanted it to do and nailed down functionality as best I could before doing design work. Started with macOS then moved to iOS. Process for building: The macOS side was pretty straightforward. Getting the grouping to work was pretty easy because I had a clear idea of how I wanted it to behave. Testing was pretty easy and iterating was quick. The iOS side was kind of nightmarish. Keep in mind I've never done this before so I was doing a lot of iterative changes with claude and the simulator and burst calling the Spotify API every time I launched. This made Spotify pretty crabby and they blocked my token for hammering for like 12 hours. Whoops. Lesson learned. I also learned that Spotify's API limits are pretty tight. If I weren't already in their system the way I am as a user I probably would have built this around something else that's more forgiving with the rate limit. I had to think about how to limit the calls but still get functionality without breaking caching rules. This is an app for 2 people to use. I get that it's their API but woof. Using the simulator: I used the simulator to do a lot of bug chasing. I don't think that was correct. It worked for some of the obvious issues but I learned that simulators are not phones so when I deployed it to my phone it had a whole host of bugs and issues that weren't able to be caught in the simulator. Also some things I thought were issues ended up resolved in the phone they were just slower in the simulator. Tracking down bugs and things that didn't work quite right: I told claude cowork that it's a project manager for finding bugs and to write prompts or briefs to help claude code solve the problems. I pointed it to the code base folder and told it to review. I did a lot of button pushing just to see what works and what didn't and fed the results back to claude cowork. It worked to get through things but is a little tedious. At one point I did catch hallucinated code on my own with imaginary endpoints claude wistfully put in there. _that wasn't easy to find._ Things that aren't bugs that require some human thought: My Sonos speakers do have limitations. Sonos answers when you ask it to do stuff. The issue is the app asks too much, too fast. (And Sonos app even goofs on this but their actual engineers seem to have smoothed it out better than me) Each tap fans out into a bunch of UPnP SOAP calls and Sonos's AVTransport coalesces overlapping ones, so 3 rapid Previous taps turn into 1 actual hop on the speaker. The work I've been doing today is mostly about asking less and asking smarter to make sure that as a user I don't accidentally make it do a metric ton of stuff when it can only really handle a few things quickly. Thing that was most fun that I didn't expect: I had a lot of fun picking out a color palate and doing the design work. I'm not artistic at all but I know what I like to look at and I'm decent at describing it. Not captu

How do you share project context with someone else so their AI is up to speed?

Curious how others handle this. When I work on a project, I usually keep a `context.md` with the background — goals, decisions, current state, open questions. My own Claude/Cursor uses it constantly. The friction starts when I want to bring someone else in — a cofounder, a freelancer, an advisor — and I want their AI to also have that context, not just them. Right now I literally: - send them the `.md` file in Telegram/Slack - a week later it's stale, so I send a new one - if I update something today, they have no idea - sometimes I just paste 5 paragraphs into a chat I know "just use a GitHub gist / repo" is the obvious answer, and for some flows it works. But it doesn't feel right when the recipient isn't a dev, or when the context evolves daily, or when I just want a clean link that their AI can fetch and that I can revoke later. Questions for the AI-heavy folks here: Do you actually run into this, or am I overcomplicating it? What do you do today? Gist? Notion share? Just paste it in chat? Has anything actually felt good? Not building anything (yet), just trying to figure out if this is a real shared pain or just my workflow being weird. submitted by /u/OsipovMe [link] [comments]